This dataset contains video frames captured from Pokémon Legends: Z-A.
Underlying audio-visual content is © Nintendo / The Pokémon Company / Game
Freak / Creatures Inc. The frames are not licensed by, nor affiliated with,
the rights holders. Only the controller-action recordings are an
independent contribution.

By requesting access you confirm that:

You will use the data only for non-commercial academic / personal
research (e.g. world-model / imitation-learning experiments).
You will not redistribute the frames, mirror them on other services,
or republish them as part of model weights that recover them.
You will respect any DMCA takedown that may be issued by the rights
holders, and acknowledge that this dataset may be removed at any time.
You are not affiliated with the rights holders in a manner that would
create a conflict of interest with this access.

arelius/nxml-pokemon-legends-za

⚠️ Underlying content © Nintendo / The Pokémon Company / Game Freak / Creatures Inc. This dataset is offered for non-commercial academic research only, under a custom research license (see LICENSE). It is not affiliated with, endorsed by, or licensed by the rights holders. If you have any commercial intent, do not use this dataset.

Behavior-cloning / imitation-learning episodes from Pokémon Legends: Z-A recorded via a Nintendo Switch capture pipeline. Each episode contains a synchronised stream of RGB frames and 26-dim controller actions (one action per frame).

Sibling repo: arelius/nxml-pokemon-legends-za-mp4 — Browser-previewable H.264/MP4 mirror.

At a glance


Episodes	234
Frames	641,039
FPS	30.0
Action dim	26 (`switch_packets.v1`)
Video	ffv1 in mkv
Resolution	480×640 (most episodes), 720×1280 (a small number)

Episodes per source capture

Source folder	Episodes
`behavior-cloning`	125
`e2e`	2
`gman`	30
`gman-battles`	3
`original`	27
`recorded-episodes`	44
`za-stages`	3

Layout

meta/
  info.json              dataset-level schema & totals
  episodes.parquet       one row per episode (episode_index, source_folder, frame_count, paths, ...)
data/
  chunk-000.parquet      per-frame rows: episode_index, frame_idx, timestamp, action
videos/
  chunk-000/
    episode_000000.mkv
    ...
    episode_000233.mkv

Episodes are renumbered episode_000000..episode_000233 for the chunked layout. The original capture filename and source folder are preserved in meta/episodes.parquet (columns episode_id and source_folder), so provenance is recoverable.

Action layout (26-dim, float32)

Order matters — existing checkpoints depend on it. Source of truth: nx_packets.action_spec (see packages/nx-packets/src/nx_packets/action_spec.py).

idx	name	type	range / notes
0	`L_STICK_X`	continuous	`[-1, 1]`
1	`L_STICK_Y`	continuous	`[-1, 1]`
2	`R_STICK_X`	continuous	`[-1, 1]`
3	`R_STICK_Y`	continuous	`[-1, 1]`
4	`L_STICK_PRESSED`	binary	0/1 (or sigmoid prob; threshold at 0.5)
5	`R_STICK_PRESSED`	binary
6	`DPAD_UP`	binary
7	`DPAD_LEFT`	binary
8	`DPAD_RIGHT`	binary
9	`DPAD_DOWN`	binary
10	`L`	binary
11	`ZL`	binary
12	`R`	binary
13	`ZR`	binary
14	`JCL_SR`	binary
15	`JCL_SL`	binary
16	`JCR_SR`	binary
17	`JCR_SL`	binary
18	`PLUS`	binary
19	`MINUS`	binary
20	`HOME`	binary
21	`CAPTURE`	binary
22	`Y`	binary
23	`X`	binary
24	`B`	binary
25	`A`	binary

Sticks are stored on-wire as int8-ish in [-100, 100]; the dataset stores them rescaled to [-1, 1]. Buttons are continuous [0, 1] (post-sigmoid for policies, or 0/1 for ground truth) — threshold at 0.5 to recover binary state.

Download

Full dataset:

huggingface-cli download arelius/nxml-pokemon-legends-za \
  --repo-type dataset --local-dir ./nxml-pokemon-legends-za

Or in Python:

from huggingface_hub import snapshot_download
snapshot_download(
    "arelius/nxml-pokemon-legends-za",
    repo_type="dataset",
    local_dir="./nxml-pokemon-legends-za",
)

Metadata only (~few MB):

import pandas as pd
from huggingface_hub import hf_hub_download

p = hf_hub_download("arelius/nxml-pokemon-legends-za",
                    "meta/episodes.parquet", repo_type="dataset")
episodes = pd.read_parquet(p)

Reading an episode

import pyarrow.parquet as pq

frames = pq.read_table("data/chunk-000.parquet").to_pandas()
ep0 = frames[frames.episode_index == 0]
# action is a length-26 list<float32> per row; cast to numpy when needed
# video at videos/chunk-000/episode_000000.mkv — decode with ffmpeg/torchvision

Codec note

Video is FFV1 lossless in MKV. Most browsers cannot preview FFV1 — download and decode locally. Designed to be bit-exact for retraining. For a browser-friendly mirror, see arelius/nxml-pokemon-legends-za-mp4.

Limitations & provenance

One source episode (gman-battles/20260213_211015_4-10pm.pt, ~9.9 GB truncated FFV1 zip) was unrecoverable and excluded.
A separate Shape-B latent-only set (post-VAE encoded .npz/.pt) was not uploaded — it is regenerable from the videos via the project's encoder.
A separate Shape-C jpg+json per-frame set was not uploaded for the same reason.
Timestamps in data/chunk-000.parquet are synthesized from a nominal FPS (30.0), not original wall-clock; the per-episode manifest's source.timestamps = "synthesized" flags this.

Rights & attribution

Video frames depict gameplay from Pokémon Legends: Z-A. The audio-visual content of the game — characters, models, environments, sound — is © Nintendo / The Pokémon Company / Game Freak / Creatures Inc. This dataset is not licensed by, endorsed by, or affiliated with any of those rights holders.
The controller-action traces (the 26-dim action vectors and the capture pipeline that produced them) are the original contribution of the uploader and are released under the terms in LICENSE (non-commercial research use only).
The licence in LICENSE does not purport to grant rights to the underlying game content. Your use of that content is at your own risk and must comply with the rights holders' terms.
If you represent a rights holder and want this dataset removed, please contact the uploader via the HuggingFace repo and it will be taken down promptly.

Citation (informal)

@misc{nxml_pokemon_legends_za,
  title  = {nxml-pokemon-legends-za: Switch-capture imitation-learning episodes (research-only)},
  author = {Saben, Clark},
  year   = {2026},
  url    = {https://huggingface.co/datasets/arelius/nxml-pokemon-legends-za},
  note   = {Frames depict Pokémon Legends: Z-A, © Nintendo / The Pokémon Company / Game Freak / Creatures Inc. Non-commercial research use only.}
}

Downloads last month: 35

Total file size:

207 GB